Graph Neural Networks for BLE Mesh

optimization

Oleksandr Bratus

Ukrainian Catholic University, Lviv, Ukraine

bratus@ucu.edu.ua

Abstract. The Bluetooth Low Energy (BLE) Mesh network technology

is one of the newest technology in the wireless communication domain.
Due to low cost and low power consumption, it has already become
widespread and has the potential for a wide range of applications. How-
ever, some disadvantages of the flooding algorithm and the problem of
interference primarily determine the task of optimizing the BLE Mesh
network. Although more and more research is devoted to BLE Mesh
optimization, applying machine learning methods to solve this problem
remains open. In addition, in recent years, the natural graph structure
of wireless interference patterns was used to define graph neural network
(GNN) architecture, therefore, GNNs have all the prerequisites for the
successful solution of the proposed problem. The main idea of this work
is to propose GNN model, which can optimize BLE Mesh network, and
test this model on simulated data.

Keywords: Bluetooth Low Energy · Bluetooth Mesh · Interference ·

Deep Learning · Graph Neural Network.

1 Introduction
In recent years, much attention has been paid to wireless communication net-
works. Along with a wide range of wireless network technologies such as Blue-
tooth, Wi-Fi, Near-Field Communication (NFC), mobile networks (e.g. 3G, 4G
and 5G), Bluetooth Low Energy (BLE) technology has become one of the most
widely-adopted technologies [4] due to high energy efficiency and widespread
availability in user equipment.
BLE technology is suitable for small portable and autonomous devices [12].
For example, it makes possible to coordinate real-time networks of wearable
devices, health care or home automation networks just by smartphone since
almost every smartphone is equipped with this technology.
The BLE provides communication between directly connected devices located
close to each other within the distance typically limited up to 50 meters, so it is
used for short-range communication. However, this is one of the main limitations
of BLE use. Including, for this reason, the technology fails to realize its full
potential, despite the widespread. Therefore, in July 2017 was released Bluetooth
Mesh [3], that wholly based on the BLE protocol stack. BLE Mesh networks
180 Oleksandr Bratus

can connect thousands of devices providing transmission of only short messages

(some bytes) between nodes with relatively small energy consumption.
Although the past few years after release Bluetooth Mesh becomes more
and more popular, few studies have been published to improve and optimise
this technology. In particular, methods are being developed to speed up the
transmission of messages and avoid message losses.
Given the topological complexity of real networks, it seems promising to
use machine learning algorithms to optimize them. In particular, Graph Neu-
ral Networks (GNNs) [14], which have been explosively evolving in recent years
and attracting the scientific community and industry’s attention, have signifi-
cant potential. The recent achievements of GNN methods show that algorithms
can outperform existing approaches, including wireless communication networks
domain. So the proposed research aims to examine, whether the latest GNN ap-
proaches able to compete with the existing methods, including complementing
them, in BLE Mesh optimization domain.
The positional paper is organized in the following order. In section II, I
present the necessary background covering the BLE mesh network standards’
main issues and give brief introduction to GNNs. The analysis of related works
is presented in Section III. Problem Setting is presented in Section IV. Section V
proposes an approach to optimize BLE Mesh network using power control meth-
ods based on Graph Neural Network methods. The research plan is presented in
Section VI. Finally, Section VII concludes this paper.

2 Background

Bluetooth Mesh uses BLE protocol stack according to which a node may possess
different features that allow each node to play notable roles in the network. There
are four types of nodes in BLE Mesh networks [13]: proxy nodes, friend nodes, low
power nodes, relay nodes. Proxy nodes can receive messages from a non-mesh-
supported BLE device and retransmit them over the BLE mesh network. The
primary function of friendly nodes in storing and transmitting messages to their
corresponding low power nodes. Friendship mechanism between these nodes is
one of BLE Mesh Standard’s latest innovations, which allows the low power node
to be asleep, conserving energy, and being activated in a case when requested by a
friendly node. Relay nodes can receive and retransmit Bluetooth Mesh messages
– they provide connectivity between non-directly connected nodes. This is one
of the most important features since the relay process makes Bluetooth mesh
multi-hop communication possible.
BLE Mesh can be seen as an application extension of the BLE stack as it
is built on top of it, inheriting Bluetooth-advantages such as robustness, low-
power and low-cost. On the other side, as BLE Mesh used in Internet-of-Things
(IoT) systems [2], it should guarantee a low energy consumption enabling a long
lifetime of battery-powered nodes. For this purpose, BLE Mesh uses simplified
mechanisms and algorithms.
 Graph Neural Networks for BLE Mesh optimization 181

Bluetooth mesh is designed to be used without the need to establish connec-

tions among devices in the network. To achieve this connection-less approach,
all data transmissions in a Bluetooth Mesh network are broadcast under the
advertising/scanning scheme of BLE [10]. Basically, BLE Mesh uses only three
broadcasting channels (2402, 2426, 2480 MHz). Each broadcast transmission is
sent using those three channels one after another during an advertisement event
to have a higher probability of messages reaching a receiver. The scanner nodes
scan a channel during a time window, and then switch into another channel - it
means they do not scan a channel all the time. BLE nodes are not synchronized
to each other, so each node has random access to the channels. This avoids en-
ergy losses when the node is in a state of waiting for a long time. However, it is
worth noting that such behaviour of nodes can cause frequent collisions, leading
to the overlap of messages, hence losing information.
The Bluetooth Mesh transmission model is based on managed flooding [9],
which defines by Mesh Standard. By this model, every node within the radio
range of transmitter receives all messages. If a receiver is a relay node, it retrans-
mits messages to nodes more distant to the transmitter. Considering that this
model does not require complex routing tables or using any routing protocols,
it reduces the required memory amount and defines the model as the simplest
among any other protocol in the IoT context. The use of flooding algorithm re-
sults in a highly resilient and reliable network due to messages’ movement to their
destination via multiple paths through the networks. The uncontrolled flow [16]
of resenting messages through the relay nodes is one of the main disadvantages
of the managed flooding algorithm and causes network bandwidth to reduce,
increases network load and collision probability. In this case, the unlimited and
uncontrolled relaying minimizes by two main methods:

– Time-To-Live (TTL) limitation sets the maximum number of hops over

which message is relayed before arriving at its destination node.
– Message cache contains all of the recently seen messages. The node compares
the newly received message with the already processed ones and decides
to ignore further or forward it. This method reduces message forwarding
without reducing network capacity and reliability.

The given basic structure allows delving into the optimization approaches
directed mainly on the optimization both topology configuration and parameters
of BLE protocol stack.

3 Related Work
My research is related to three separate topics. First is the study of optimal
parameter settings of BLE Mesh network. Both the correct optimized messaging
protocol settings and the optimized topological configuration can significantly
improve network performance. Second is the study of the influence of interference
on the reliability of wireless communication networks. Understanding the mech-
anism of interference will mathematically formalize the interference process to
182 Oleksandr Bratus

assess its impact on the transmission of messages. The last is Graph Neural Net-
works’ use to optimize the transmission of messages in wireless communication
systems.

3.1 Classical approaches for BLE Mesh optimization

Since BLE Mesh networks’ work is based on the Bluetooth Mesh protocol, much
of the works are devoted to analyzing the configuration of all the different pro-
tocol’s parameters.
Paper [13] provides general guidelines to ensure the reliability and scalabil-
ity of the network. It focused on the packet broadcast process structure, whose
performance depends mainly on time parameters such as duration of scanning
and advertising intervals. Such network performance indicators guide the au-
thors as packet loss rate, the time delay, and packet errors per unit time in
their study. They conclude that the correct operation of the network protocol
strongly depends on each of the time parameters’ mutual settings, so the appro-
priate setting of both advertising and scanning processes has a decisive influence
on transmitting messages. However, experiments have shown that scalability re-
mains problematic for these BLE Mesh networks due to message flooding what
leads to the broadcasting storm and, as a result, reduced network reliability.
Paper [9] devoted to the study of the possibilities and limitations of the
Bluetooth Mesh protocol. All experiments are performed on a static network
with fixed time parameters. The paper proposes two metrics to measure BLE
Mesh protocol performance and capacity to deliver messages reliably. End-to-
end Packet Delivery Ratio (PDR) calculated as the proportion of successfully
sent packets by a particular node (which reached the destination node) for a
time window. Burst Drops (BD) represents the number of packets losses in a
row. Authors are interested in observing the maximum BD for different nodes in
experiments as this indicates extreme cases where the nodes lose connection and
can cause the network to crash. As a result, the paper describes the data rate
limitations of the technology. It is concluded that when nodes are oversaturated
and transmitting messages with a high frequency, the technology cannot provide
an acceptable packet delivery performance.
Another way to improve BLE performance is to maximize the lifetime of the
devices [10]. Authors present an extension and optimization of the more com-
plete and precise model proposed in [8]. They propose an algorithm based on the
configuration of the three time-parameters of the BLE protocol (scanning inter-
val, scanning window and advertising interval) to minimize power consumption.
Experimental results indicate a significant increase in the lifetime of the devices
dozens of times.
As mentioned earlier, managed flooding can cause a large number of colli-
sions. This uncontrolled situation is called a broadcasting storm characterized
by a significant decrease in a packet delivery ratio. The next paper [7] is pro-
posed on relay selection approaches, where only a few nodes are relays and can
retransmit messages. Authors proposed three heuristic algorithms whose task to
form a connected set of relays so that each network node has been connected
 Graph Neural Networks for BLE Mesh optimization 183

to at least one relay node. The algorithms evaluated are Greedy Connect, K2
Pruning and Dominator!. Each algorithm outputs a fully connected network and
can run distributed on each network node without requiring centralized control.
To compare the performance of algorithms is used Packet Delivery Ratio (PDR)
metric. Authors conclude that each of the algorithms increases packet delivery
probability experiencing the conditions of optimal use of each of them.
One more paper [2] describes a relay node management method. To minimize
the number of active relay nodes of the BLE Mesh network, authors formulate
the Minimum Relay Tree (MRT) problem and provide an exact solution based
on integer linear programming. The authors emphasize the efficiency of the algo-
rithm in dense networks. They conclude method reaches a compromise between
the amount of energy consumed and message delivery ratio.

3.2 Interference problem

While introducing a new wireless communication network, it is crucial to ensure

that it will not interfere with the existing networks. Paper [17] describes a model
that able to model interference conflict between different radio systems. The
method’s main achievement is the ability to model both the victim signal and
the interfering signal taking into account such variables as modulation, power,
frequencies, losses, filters, and many others. Also, the model uses information
about differences in distance between systems and their difference in frequencies.
Another method presented in [6] uses a supervised-learning classifier to detect
interference already at the level of single interference bursts. The authors claim
that the method can isolate and extract standard-specific traffic statistics even
in the case of heavy concurrent interference. Results of experiments demonstrate
up to 97% interference identification accuracy.
One of the possible approaches [13] that consider the behaviour of a Bluetooth
Mesh network under the influence of interference from WLAN uses existing
signal-to-interference-plus-noise-ratio (SINR) [12].

PRx
SIN R = , (1)
N + IW + IBM
where PRx is the power of the signal received by a scanning device, N is noise
floor, IW is power of the WLAN interference, IBM is power of the interference
from other BLE Mesh devices transmitting on the same channel. Then Packets
Error Ratio (PER) for an Additive White Gaussian Noise (AWGN) channel can
be calculated as a function of SINR [4].
 p n
P ER = 0, 5 · erf c 0, 68 · SIN R , (2)

where erf c(·) is the complimentary error function, n - size of the packet in bits.
Such a method can detect real-time interference; it can identify and map
external interference from different wireless systems, according to their power
and temporal features.
184 Oleksandr Bratus

3.3 GNN approaches to wireless power control

Recently, number of applications where data are generated from non-Euclidean

domains is dramatically increased. It turned out that traditional methods of ma-
chine learning are not able to effectively process such data. Therefore, models
have been proposed that are based on existing deep learning approaches, but
accept inputs in the form of graphs with complex relationships and interdepen-
dency between objects. These methods are called Graph neural networks and are
divided on Recurrent graph neural networks (RecGNNs), Convolutional graph
neural networks (ConvGNNs), Graph autoencoders (GAEs), Spatial-temporal
graph neural networks (STGNNs), which follow the appropriate deep learning
algorithms [18].
Although GNNs have become widely known and rapidly evolving, under-
standing of their capabilities and limitations remains poorly understood [19].
Most studies are based on the empirical results of narrowly focused areas, using
GNNs without a clear rationale for the approach.
One of the areas where the GNNs have been applied is a reflection of complex
wireless communication systems with nontrivial topological structure. Paper [15]
dedicated to solving the power control problem in K-user interference channels.
Authors developed an Interference Graph Convolutional Networks (IGCNet),
the main property of which is to capture the permutation invariance property
of the interference channel. Such network is also robust the inaccurate measure-
ments, which is especially important in a real environment where every message
is subject to perturbation and interference. The advantages of IGCNet include
scalability and ability to solve weighted problems (power control problem in
particular).
Paper [11] considers a similar problem, but uses other assumptions and ap-
proaches. Authors present the Random Edge Graph Neural Network (REGNN)
architecture as a parameterization for the resource management policy. To re-
duce the impact of interference on concurrent transmissions authors propose
to use use an unsupervised optimization approach to learn optimal power allo-
cation decisions. Such approach allows to restrict a minimum rate constraint,
which adapts to the network size. These means model achieves the right balance
in power allocations throughout a range of network configurations.
Based on these results of both papers, we can assume that described GNN
models can have a great influence on the solution of our optimization problem. It
is worth noting, every node in BLE Mesh network can change signal transmission
power and it is possible to build model to optimize power of nodes as at too strong
powers signals will be superimposed on each other, and at too little powers signals
will not reach the nearest nodes on message transmission.

4 Problem Setting

The main objective of this project is to propose an approach to optimize BLE

Mesh network using GNN methods, such as REGNN and IGCNet, which have
 Graph Neural Networks for BLE Mesh optimization 185

already shown good results in wireless power control domain and claim to be ap-
plied to a wide range of tasks optimization of wireless communication networks.
Besides that the following goals are aimed:

• adapt GNN methods under restrictions of BLE Mesh networks;

• evaluate the performance both algorithms (on the simulated data);
• compare the performance both algorithms with performance classical non-
machine learning algorithms introduced earlier.

5 Approach

5.1 Data simulator

To achieve our objectives, we need develop a BLE Mesh network simulator that
would check models outputs (network configuration). The basic principles of the
simulator should be based on the BLE Mesh Standard and on BLE protocol to
correctly display topological features of the mesh network and adequately simu-
late the process of messaging between different types of nodes. In this problem’s
context, it seems appropriate to take into account the article [1] that studies
various mesh topology formation techniques and focuses on how mesh networks
can be established over BLE communications. In addition, the simulator must
take into account the effects of interference, which was discussed in detail in
previous sections. Existing BLE Mesh network implementations 1 should also be
considered.

5.2 Model pipeline

Though baseline models do not give a good result, it always is better to have
such a model. In the early stages, baseline models can help us to notice several
issues, which can be noticed late in the more advanced model. Furthermore, it
would be simpler to test some experiments quickly using the baseline instead of
the advanced approach.
As a baseline model I propose to take an heuristic on relay selection algo-
rithm which is guaranteed to optimize our network, and will allow to test the
correctness of the simulated data.
Adaptation of the such advanced models as IGCNet and REGNN to research
domain is primarily to select the right objective and loss functions which should
take into account all the restrictions of BLE Mesh networks. In case of the power
control problem in K-user interference channels [5] the objective function to be
maximized is defined as follows:
K
X
f (p) = wk log2 (1 + SIN Rk ),
k=1

1
https://github.com/ogoremeni/bluemesh7
186 Oleksandr Bratus

0 ≤ pk ≤ Pmax , ∀k,
where p = [p1 , ..., pK ] denotes power allocation vector, wk is the weight for
the k th pair. Then the loss function is defined as the negative sum rate:
K
X
l = −EH wk log2 (1 + SIN Rk (θ)) ,
k=1

where θ denotes the parameters of the neural network.

Based on a formal definition of the power control problem in K-user interfer-
ence channels, it is advisable to limit the power limits of nodes that are not relays.
In addition, the power of the relay can be adjusted according to the number and
types of nodes surrounding the relay, i.e. according to the network density at the
corresponding point. Such restrictions seem to clearly classify relays and non-
relays and prevent unwanted power distributions in advance. Another important
addition is to determine the relationship between the relays. One possible option
is to give the cross-link channels hjk and hkj connecting the relay j and k more
weight, because the flow of messages between them is expected to be particularly
high. Such an innovation can significantly increase the throughput of relays and
avoid delays and loss of messages.
If objective and loss functions are selected successfully (including justification
for its choice), I assume that the convergence of the loss function to its optimum
will directly mean the optimization of the input BLE Mesh network.

5.3 Evaluation

There also should be provided evaluation result that shows how well proposed
models optimize BLE Mesh network. The chosen main metrics are the following:

(1) average sum-capacity

(2) Packet Error Rate (PER)

(1) used to compare the outputs of graph neural networks in solving power
control problem.
(2) is the universal performance metric, since different models may have
different loss functions or different parameters to optimize. This metric is the
most representative, as it indicates both network reliability and sensitivity to
changes network parameters.

6 Research Plan

November 2020

1. Coordination of the direction of the thesis

2. Formulate a rough scope of research and the main objectives
3. Review of past and current related work
 Graph Neural Networks for BLE Mesh optimization 187

4. Write and submit abstract to MS-AMLV-2021

December 2020
1. Formulate refined research objectives
2. Look for GNN algorithms implementations that can be appropriate in our
research domain
3. Write and submit a proposal paper to MS-AMLV-2021
January 2021
1. Look for implementations of BLE Mesh network simulator
February 2021
1. Implement BLE Mesh network simulator
2. Make sure simulated data imitate behavior of the real-world data
3. Refine the final version of the position paper to MS-AMLV-2021, submit and
prepare an oral report
March 2021
1. Implement and evaluate on simulated data classical relay node management
methods
2. Analyze the evaluation results and write conclusions
3. Adapt pre-trained GNNs to achieve our goals
4. Analyze the evaluation results and write conclusions
April 2021
1. In case of a success GNN approach, identify the main advantages of the
approach and indicate the limitations
2. Compare results to the State-of-the-Art methods for BLE Mesh network
optimization, formulate conclusions
3. Working on the diploma text
May 2021
1. Write the final sections to the Master thesis paper and refine the whole work
2. Submission of thesis for final review

7 Conclusion
In this master’s thesis proposal, I made an overview of the current state in the
field of BLuetooth Low Energy Mesh network optimization and described graph
neural networks, which effectively captures hidden patterns of graph-based data
and hence applicable in proposed research domain. After literature review, where
I identified the potential problems of existing approaches, I assume that GNN
methods have great potential to outperform those approaches. Finally, I stated
clear and achievable project goals, proposed the approach to achieve those goals
and briefly described plan of the project work.
188 Oleksandr Bratus

References
1. Agnihotri, M.K.: Energy efficient topology formation for bluetooth mesh networks
using heterogeneous devices (2015)
2. Beben, A., Bak, A., Sosnowski, M.: Efficient relay node management in ble mesh
networks. International Journal of Electronics and Telecommunications 66(1), 29–
35 (2020)
3. Bluetooth SIG: Bluetooth mesh profile specification v1.0 (2017)
4. Chang, K.H.: Bluetooth: a viable solution for iot? [industry perspectives]. IEEE
Wireless Communications 21(6), 6–7 (2014)
5. Eisen, M., Zhang, C., Chamon, L.F.O., Lee, D.D., Ribeiro, A.: Learning optimal
resource allocations in wireless systems. IEEE Transactions on Signal Processing
67(10), 2775–2790 (2019)
6. Grimaldi, S., Mahmood, A., Gidlund, M.: Real-time interference identification via
supervised learning: Embedding coexistence awareness in iot devices. IEEE Access
7, 835–850 (2019)
7. Hansen, E.A.J., Nielsen, M.H., Serup, D.E., Williams, R.J., Madsen, T.K., Abild-
gren, R.: On relay selection approaches in bluetooth mesh networks. In: 2018 10th
International Congress on Ultra Modern Telecommunications and Control Systems
and Workshops (ICUMT). p. 8631214 (2018)
8. Kindt, P., Yunge, D., Diemer, R., Chakraborty, S.: Precise energy modeling for
the bluetooth low energy protocol. arXiv preprint arXiv:1403.2919 (2014)
9. Leon, E.D., Nabi, M.: An experimental performance evaluation of bluetooth mesh
technology for monitoring applications. In: 2020 IEEE Wireless Communications
and Networking Conference (WCNC). pp. 1–6 (2020)
10. Liendo, A., Morche, D., Guizzetti, R., Rousseau, F.: Ble parameter optimization
for iot applications. In: 2018 IEEE International Conference on Communications
(ICC). pp. 1–7 (2018)
11. Naderializadeh, N., Eisen, M., Ribeiro, A.: Wireless power control via counter-
factual optimization of graph neural networks. In: 2020 IEEE 21st International
Workshop on Signal Processing Advances in Wireless Communications (SPAWC).
pp. 1–5 (2020)
12. Raza, S., Misra, P., He, Z., Voigt, T.: Bluetooth smart: An enabling technology for
the internet of things. In: 2015 IEEE 11th International Conference on Wireless
and Mobile Computing, Networking and Communications (WiMob). pp. 155–162
(2015)
13. Rondon, R., Mahmood, A., Grimaldi, S., Gidlund, M.: Understanding the perfor-
mance of bluetooth mesh: Reliability, delay, and scalability analysis. IEEE Internet
of Things Journal 7(3), 2089–2101 (2020)
14. Scarselli, F., Gori, M., Tsoi, A.C., Hagenbuchner, M., Monfardini, G.: The graph
neural network model. IEEE Transactions on Neural Networks 20(1), 61–80 (2009)
15. Shen, Y., Shi, Y., Zhang, J., Letaief, K.B.: A graph neural network approach for
scalable wireless power control. In: 2019 IEEE Globecom Workshops (GC Wkshps).
pp. 1–6 (2019)
16. Tseng, Y.C., Ni, S.Y., Chen, Y.S., Sheu, J.P.: The broadcast storm problem in a
mobile ad hoc network. Wireless Networks 8(2), 153–167 (2002)
17. Vihlborg, F.: On models for interference calculations between radio communication
systems (2011)
18. Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., Yu, P.S.: A comprehensive survey
on graph neural networks. IEEE Transactions on Neural Networks pp. 1–21 (2020)
 Graph Neural Networks for BLE Mesh optimization 189

19. Xu, K., Hu, W., Leskovec, J., Jegelka, S.: How powerful are graph neural networks.
In: International Conference on Learning Representations (2018)